Abstract
Research on improving the performance of feedforward neural networks has concentrated mostly on the optimal setting of initial weights and learning parameters, sophisticated optimization techniques, architecture optimization, and adaptive activation functions. An alternative approach is presented in this paper where the neural network dynamically selects training patterns from a candidate training set during training, using the network's current attained knowledge about the target concept. Sensitivity analysis of the neural network output with respect to small input perturbations is used to quantify the informativeness of candidate patterns. Only the most informative patterns, which are those patterns closest to decision boundaries, are selected for training. Experimental results show a significant reduction in the training set size, without negatively influencing generalization performance and convergence characteristics. This approach to selective learning is then compared to an alternative where informativeness is measured as the magnitude in prediction error.